Clustering single-cell RNA-seq data by rank constrained similarity learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell RNA sequencing (scRNA-seq) enables identification of heterogeneous cell types.
• scRNA-seq data presents challenges due to biological noise, technical artifacts, and high dimensionality.
• Existing tools for cell type identification require accuracy improvements.

Purpose of the Study:

• To introduce a novel clustering algorithm and tool, RCSL (Rank Constrained Similarity Learning), for accurate cell type identification from scRNA-seq data.
• To address the limitations of existing methods in handling complex tissue data.

Main Methods:

• RCSL integrates local and global similarity measures for cells.
• Global similarity is assessed using Spearman's rank correlations of expression vectors.
• Local similarity is determined by adaptively learned neighbor representations.
• The algorithm automatically estimates cell type numbers and identifies clusters by minimizing distance to a block-diagonal matrix.

Main Results:

• RCSL demonstrated superior accuracy and robustness compared to six state-of-the-art methods on 16 benchmark scRNA-seq datasets.
• The algorithm effectively discerns subtle differences within cell types and larger differences between cell types.
• Performance was validated using established metrics on well-annotated datasets.

Conclusions:

• RCSL offers a significant advancement in cell type identification from scRNA-seq data.
• The method provides a robust and accurate approach for analyzing complex biological samples.
• RCSL is available as an R package for broader research application.